Gate passes 12 and 13 rejected the prefix-grouped weblate_network_*.py modules, the stand_ins.py test helper and unmirrored test paths. The modules are now weblate_network/strategy.py and middleware.py, mounted as a namespace package under /app/infinito and referenced by their dotted package names; the helper is _stand_ins.py and the module tests mirror the package.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Weblate built every absolute URL, the social-auth redirect and the CSRF origin from WEBLATE_SITE_DOMAIN, so the onion sibling either redirected to clearnet or failed the OIDC round trip. A settings override swaps in a network-aware social strategy and proxy middleware that take scheme and host from the request, drops the forced HTTPS redirect the http onion vhost cannot satisfy, and ALLOWED_HOSTS now lists every network sibling plus the container name and loopback. The two modules are separate because social_django loads the strategy at app-ready before defining load_strategy, which weblate.middleware imports back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cadvisor ran with its defaults: every series of every cgroup carried the union of all containers' label keys (97 on echoserver) and every cgroup was re-read each second, so each scrape grew with containers times images and the 0.25-CPU cadvisor was OOM-killed in its 1g cgroup about every 13 minutes. It now stores only the compose and swarm identity labels and keeps housekeeping at 10s, below the 30s scrape interval; limits and metric groups stay unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Gate pass 9 on the merged store-i18n stopped at ruff N818 on the local Stop exception that the builder-timeout test defined next to the module's StopBuilderError from the build-before-clone tests, plus SIM117 on the nested with in the healthy-mirror test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fetch() leaked subprocess.TimeoutExpired, which ended the builder thread while it kept the builder lock, so no replica built again. A clone killed at its timeout also left a HEAD that never resolves, and the HEAD-file check ran remote update on it forever. The loop and refs() now catch SubprocessError, fetch() re-clones when HEAD does not resolve, and the clone lands in repo.git.partial and replaces the mirror only after it succeeds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On real GitHub Actions INFINITO_BUILD passed through from default.env as 1, so every deploy job built the infinito image locally (~5 min) and then pulled and ran the published ci image anyway (pull_policy always). An env handler now defaults INFINITO_BUILD to 0 there; act and jobs that set it themselves keep building.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
run_builder fetched the git mirror before it claimed any build, so the deployed site, which copies the snapshot baked into the image and reads no mirror, waited for the first git clone --mirror (about 60-150 s on CI, 967 MiB with 1113 refs/pull/* heads). The builder now claims and submits DEPLOYED inside the pool before the first fetch; latest and the tags still wait for it, so none builds from an outdated head. The README no longer claims the fetch waits for a running build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
test_used_keys_declared reads every INFINITO_* literal in code as an environment key, so the marker string copied from vars/main.yml failed the lint without a registry entry. The marker only lands in the append's header comment, which the test does not inspect.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
taiga-back's image settings build STATIC_URL from TAIGA_URL, the canonical clearnet domain, so the Django admin on the onion vhost linked base.css and its scripts to the clearnet host and the onion CSP blocked them (header background rgba(0, 0, 0, 0) in the design spec). STATIC_URL is now the same-origin path below FORCE_SCRIPT_NAME. The Jinja static() global of the e-mail engine keeps the canonical absolute URL, because a mail client resolves no same-origin path.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The access-log map read $request_uri, which keeps percent-encoding and
repeated slashes. A request such as /%70assword/reset/<token>,
/password%2Freset/<token> or //password//reset/<token> missed the
token-route expression, fell through to the plain-path branch and was
logged with the token in plain text, while the backend still decoded it
to its reset handler.
The map now reads $uri: decoded, slash-merged (merge_slashes is not
overridden anywhere) and without the query string, so the existing
expressions match every spelling of a token route. Trade-off: after an
internal redirect the log shows the internal URI; with the generic
vhost's error_page (ws_generic.conf.j2) a 5xx line now carries
/500.html instead of the client's path, with the status unchanged.
Nothing in the repository parses the proxy access log.
The unit test resolves the map against the variable it reads, modelling
$uri as nginx builds it, and adds the three encoded spellings above.
Negative control: with the map back on $request_uri, exactly those three
cases fail (3 failed, 8 passed).
Reported by CodeRabbit on #1322.
Part of #1192
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pins of every role applied at every round index, so a role the round did not deploy could pull a partner into a heavier variant (web-app-discourse#1 measured 35.46 GB against the 32 GB reservation cap). resolve_round now starts every role on its clamped index and recomputes the include and the pins in force until the assignment settles, raising VariantPinError on a cycle. The planner resolves each round once (39.1 s to 16.0 s for all rounds), affected judges every (role, index) with the round the planner deploys for it, and the resource budget lint measures the pin-aware rounds. The stale web-app-matomo pin above the primary-only docs entry is dropped, and test_variant_pin_fixed_point.py guards every planned round against unsettled pins and pins above a disabled entry. docs/contributing/artefact/files/role/variants.md explains the settle loop with diagrams.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The server answered 202 for every /deployed/<code>/ path until the deployed Sphinx build finished (about 80 min with -j1), so test-languages.js:61 outwaited its poll budget on attempt one and passed only on retry. The untranslated decision now reads the snapshot's meta/languages.yml and docs catalogs, the same files the build writes into the language index, so the 404 set is identical before and after the build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dedicated deploy launched whatever ansible-playbook PATH found. A stale pkgmgr venv without Pillow crashed the galaxyserver deploy after 4h23m at the first branding task. The runner now verifies every distribution of the pyproject deploy extra in the running interpreter before make setup, exits naming the missing ones with the install command, and runs the ansible-playbook next to that interpreter.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A user key that no role declares, such as an inventory leftover, gets no username, firstname, lastname, uid or gid derived. openldap_user_entries then crashed with a bare KeyError 'username'. The filter now raises an AnsibleFilterError that names the user and the missing fields.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The front proxy logged the full request line: production had no
access_log, so nginx fell back to `combined` ($request, Referer,
$remote_user), and the debug format logged $request and the Referer.
Query-string tokens (?access_token=, ?key=, ?token=) and token routes
such as /password/reset/{token} and /i/confirm-email/{user}/{random}
ended up in plain text in the container log.
nginx.conf.j2 now declares a map from $request_uri to
$infinito_access_log_path:
- the path is written without the query string;
- the segments after a known token route (password/reset,
password-reset, reset-password, confirm-email,
lostpassword/reset/form, email-login, activate-account; matched
case-insensitively at any depth) are written as [redacted], so the
request, its status and timings stay visible without the token.
Both modes write an explicit access_log to /dev/stdout:
- production uses the new `redacted` format, laid out like `combined`
(remote address, time, "method path protocol", status, bytes,
user agent) with Referer and basic-auth user written as "-", so
combined-style parsers keep working;
- debug keeps its fields (host, timings, upstream time, lengths,
X-Forwarded-For, scheme, server name) but logs
"method path protocol" instead of $request and drops the Referer.
The path comes from $request_uri instead of $uri so internal rewrites
(error_page, try_files) do not replace the path the client requested.
Nothing in the repository parses the proxy access log (Wazuh, the
health checks and the Prometheus log_by_lua metrics do not read it),
so no consumer needs to change. The error log still carries
`request:` and `referrer:` on error-level entries; that is outside
this change.
A unit test renders nginx.conf.j2 in both modes and asserts that every
access_log uses a format declared there, that no format references
$request, $request_uri, $args, $query_string, $is_args, $arg_*,
$http_referer or $remote_user, and, by resolving the map as nginx does,
that query strings are dropped, token routes are redacted and
lookalike paths stay unchanged.
Verified live with web-svc-html in production mode (deploy failed=0, so
openresty accepted the config). A probe request to
/password/reset/<probe>?token=<probe> was logged as
"GET /password/reset/[redacted] HTTP/1.1" 404 62 "-" "curl/8.14.1",
and /plain/page as "GET /plain/page HTTP/1.1". Negative control: the
same request against the previous config logged the probe path token
and the query token in plain text. The debug format with the fix is
covered by the unit test only.
Closes#1192
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI run 41147980556 Nextcloud#0 failed both conductor design specs on a
multi-network node. e0179a0893 handed every conductor the other
network's conductor in LOGOUT_SIBLINGS. With that list set,
universal-logout v1.5.0 hides #iframe-warning on the stand-alone page
("'#iframe-warning' matches nothing visible"), and v1.5.1 fetches the
sibling after its sweep, which the conductor's per-network CSP blocks
("Connecting to 'https://logout.main.infinito.test/' violates ...
connect-src 'self' *.<onion>"). Requirement 037 limits each vhost's CSP
to its own network, so the browser hand-off could never complete in
either direction; its only effects were the two red specs.
Drop the LOGOUT_SIBLINGS wiring and its README bullet, so the compose
template and README match their state before e0179a0893. Each conductor
still logs out of every domain of its own network; ending the sessions
of the other network belongs on the server side, not in the browser.
The test now asserts that no conductor on a single-network, tor-only or
multi-network node is handed a conductor of another network.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A production redeploy of echoserver hung at the swap-forge task and the
host stopped answering SSH. swap-forge v1.0.1 recreates the swapfile
whenever its size differs from the requested MiB by any amount, starting
with swapoff on the live file. The request is the host's MemTotal, which
drifts by a few MiB across kernels, so a routine deploy paged the whole
swap back in, charged it to the containers' memory cgroups and starved
sshd until the OOM killer took swapoff.
Read the active swap files with swapon before swap-forge runs and skip
it when an active file is within 5 % of the requested size, reporting
the kept size against the request. The sizing formula is unchanged, so
no host is forced into a one-time recreate; a deliberate resize beyond
the tolerance still reaches swap-forge.
The unit test resolves the real vars through Ansible's Templar for MiB
drift both ways, no swap, zram and partitions, a halved size, an exact
pin, a lowercase unit and an escaped file name, and shows that the exact
MiB comparison swap-forge uses misses the drift.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fork CI run 41147980556 pgAdmin#0 failed in sys-ctl-hlth-runtime with
"stalwart_webmail: PHP 8.4.26 is below the security floor 8.4.27".
The deploy ran the fork's GHCR mirror of roundcube/roundcubemail:1.7.5-
apache, copied on 10-19 with PHP 8.4.26, while upstream re-pushed the
same tag on 10-27 with PHP 8.4.27. Branch CI runs the sync with
--only-missing, which skipped the tag because it existed ("destination
exists, skipping") without comparing digests, and the full refresh only
runs on the fork's default branch, which still pins 1.7.4.
- utils/docker/registry.py: manifest_digest() reads the
Docker-Content-Digest of a reference with a HEAD request, which does
not count against Docker Hub's pull allowance.
- GHCRProvider.is_stale() reports a mirrored tag as stale only when both
registries answer with a digest and the two differ.
- Under --only-missing an existing tag is re-copied when it is stale; a
failed re-copy is reported on stderr with the image and keeps the
existing copy, so one refresh cannot block every deploy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The shared proxy's nginx.conf derived worker_connections from the
openresty container's pids_limit. pids_limit caps processes and threads,
worker_connections caps simultaneous sockets per worker (client and
upstream), so with pids_limit 256 and one worker the whole host's proxy
held about 256 sockets, roughly 128 proxied requests in flight across
every vhost. Raising pids_limit to buy connections would loosen the
fork-bomb guard.
svc-prx-openresty now declares services.openresty.worker_connections
(4096 per worker, the common nginx production range, enough for several
chunk-heavy apps such as OpenWebUI at ~150-170 requests per page load).
WEBSERVER_PROXY_WORKER_CONNECTIONS reads it through lookup('config'),
without a default, so a missing key fails loudly. pids_limit stays 256.
WEBSERVER_PROXY_WORKER_RLIMIT_NOFILE is derived as 2x worker_connections
(8192), one descriptor for each client and upstream socket of a proxied
request, and nginx.conf renders it as worker_rlimit_nofile. The openresty
container pins its nofile ulimit (soft and hard) to the same value:
workers call setrlimit as root without CAP_SYS_RESOURCE, so they can only
lower the hard limit, and the inherited hard limit depends on the host
daemon (Docker >= 25 defaults to 1024:524288, DiD nests whatever the
outer container has). An explicit ulimit makes the rlimit succeed on
every host instead of logging a setrlimit alert.
The unit test pins worker_connections 4096, worker_rlimit_nofile 8192 and
the container ulimit, and fails when worker_connections follows
pids_limit again.
Live proof (compose deploy of web-svc-html, both plays failed=0): the
deployed nginx.conf renders worker_rlimit_nofile 8192 and
worker_connections 4096; openresty's master, worker and cache manager
run with "Max open files 8192 8192"; the openresty log carries no
setrlimit or "worker_connections are not enough" line.
Closes#1300
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
test_sso.py, test_sso_oidc_plugin.py and test_sso_proxy_consumers.py
evict utils, utils.cache, utils.cache.applications and the roles
packages from sys.modules to import their plugin against fresh modules,
and never put the originals back. Ansible's plugin loader keeps the
applications lookup it loaded earlier in sys.modules, so that lookup
stayed bound to the evicted utils.cache.applications, whose
_MERGED_APPLICATIONS_CACHE utils.cache._reset_cache_for_tests() no
longer reaches.
On main this turned test_webserver_worker_limits red whenever one xdist
worker ran a test that loads the applications lookup through Ansible,
then one of these sso tests, then the worker-limits test: the stale
cache served the cpus 0.5 payload of test_a_sub_cpu_proxy_still_runs_one_worker,
so the later methods rendered worker_processes 1 instead of 4 (main run
41139196554, Workspace centos and fedora, second make test on gw1).
_load_module now runs inside mock.patch.dict(sys.modules): the plugin
keeps the fresh modules it was built against, and sys.modules returns to
its prior state afterwards. This is the change store-i18n carries since
ec0fde6561, ported unchanged so both lines stay identical.
Evidence:
- Before: pytest -n 0 test_tor_ports.py test_sso_oidc_plugin.py
test_webserver_worker_limits.py -> "3 failed, 33 passed", the CI
failure set ('1' != '4' twice, worker_processes 1 vs 4).
- After: test_tor_ports.py, the four sso test files and
test_webserver_worker_limits.py -> "68 passed".
(cherry picked from commit ec0fde6561)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI run 40979252521 Weblate#1 (swarm) never converged: every task after
the first died in collectstatic with "PermissionError: [Errno 13]
Permission denied: '/app/cache/static/weblate_fonts'". The tmpfs at
/app/cache/static sits inside the shared weblate_cache volume; the first
task's runc creates the mountpoint as root 0755, and runc gives a tmpfs
on an existing mountpoint that directory's mode. 99a7d28067 declared
mode 1777, which compose honours, but docker stack deploy keeps only the
tmpfs size (docker/cli 29.9.2 convert/volume.go), so swarm still got a
root 0755 tmpfs.
- utils/storage/nfs.py: nested_tmpfs_mountpoints() derives every tmpfs
mountpoint whose innermost enclosing volume of the same service is
NFS-backed, with its declared mode; a nested tmpfs without a mode, or
one mountpoint declared with two modes, raises.
- plugins/lookup/nfs_tmpfs_mountpoints.py feeds it the rendered compose
volumes, so compose_volumes stays the source of which volumes live on
NFS.
- sys-svc-compose nfs_prep.yml creates those mountpoints on the NFS
server with the declared mode before every stack deploy, which also
repairs a mountpoint an earlier deploy left at 0755.
- web-app-weblate: the cache volume declares nfs uid 1000, gid 0,
mode 0770; a pre-created mountpoint makes the volume non-empty, so
Docker no longer copies the image's /app/cache root in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
main's admin allowlist test renders sys-svc-proxy's location/html.conf.j2,
which calls the entity_name Jinja filter. The test stubbed it as
get_entity_name, the name store-i18n retired when it registered the
filter as entity_name (plugins/filter/get/entity_name.py), so all eleven
cases failed with "No filter named 'entity_name'" after the main merge.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cursor#0 (CI run 40979252521, debian host) installed the pinned cursor_3.27.6 deb on pass 1; its postinst registers Cursor's apt repo, so pass 2's dist-upgrade moved cursor to 3.27.22, and the plan's ansible.builtin.apt deb=<3.27.6> then failed with 'A later version is already installed': the apt module skips only an equal version and otherwise fails or, with allow_downgrade, downgrades. package_install now asks the target before such a call: dpkg-deb --field reads Package and Version from the checksum-verified deb, dpkg-query reads the fully installed version, and dpkg --compare-versions decides; at or above the pin the call is skipped with changed false, so the deploy-time upgrade may move vendor packages forward and the pin never pulls them back. Any query that cannot be read falls through to the apt install, which decides loudly. The dnf path already skips an older local rpm, so both families now behave alike.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Main CI run 40976399687, socialhome#0 (swarm, tor, centos): the CLI tests
of svc-net-tor (onion:25 "socks CONNECT failed with reply 0x01") and
web-app-stalwart (probe_mail relay: ConnectionRefusedError on port 25)
failed. dockerd on the manager had aborted the whole ingress program of
the mail stack:
Failed to add ingress: failed to program ingress ports:
failed to bind host port 0.0.0.0:993/tcp: address already in use
/proc/net/tcp named the squatter: 192.168.244.10:993 -> :2049, the
node's own NFS client. The mount carries no noresvport, so the kernel
drew a source port from the 665-1023 reserved window and landed on
imaps. None of 25/465/993/995/4190/8095 got a DNAT rule or a
docker-proxy listener while every other published port did, and
stalwart still reported healthy because its healthcheck runs inside the
container. 993 and 995 are the only published ports inside that window,
so this is luck of the draw on any NFS-backed swarm host that runs the
mail role.
mount_opts now emits noresvport. insecure on the kernel-flavor export is
the other half: kernel nfsd defaults to secure and answers a mount from
an unprivileged source port with EACCES. It reaches the kernel flavor
only; ganesha already accepts unprivileged clients.
What it costs: the next deploy of every existing NFS client sees a
pending opts change and remounts the shared state; a host-side job
holding that path at that instant fails loudly. A plain remount may keep
a running mount on its old reserved port until the next real mount.
Port of f46c4d0bab from feature/store-i18n, applied identically so the
later store-i18n -> main merge stays conflict-free.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The shared pacman cache was the first writable CacheDir on every arch runner and lab node. Our arch image sets DisableSandbox, so libalpm downloads straight into that directory (util.c _alpm_download_dir_setup only makes a per-process temp dir with the sandbox on), and the XferCommand writes <pkg>.part there with curl -C -o %o; two nodes fetching the same package appended to the same .part file, and the second rename failed or left a corrupt package in the shared cache.
pacman.sh now gives each node its own download directory <shared>/pacman/node-XXXXXX/ (created once with mktemp, reused from pacman.conf on later runs) as the CacheDir between the read-only host layer and <shared>/pacman/, and writes a PostTransaction hook to pacman.d/hooks/infinito-shared-cache.hook that renames every finished package and signature into <shared>/pacman/; .part files stay for the next resume. Both directories sit in one bind mount, so each publish is an atomic rename(2). pacman.sh fails when pacman.conf sets HookDir, since pacman would then skip the hook silently. Requirement 065 and cache.md describe the three layers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
